Why an AI Pilot Works at One Plant and Breaks at the Second
A panel at the Industrial AI Summit 2026, reported by IIoT World on 18 November 2026, argues that operational context and data semantics, more than data cleanliness, are what fail to transfer between plants. The content is sponsored and the speakers are vendors, but the practical points are useful for anyone working on PLC, SCADA and integration.
On 18 November 2026, IIoT World published an article, labeled as sponsored by Arch Systems, on a problem familiar to anyone who works on the plant floor: an AI pilot project succeeds at one plant and then stalls at the second. The article grew out of a panel at the Industrial AI Summit 2026 featuring Tim Burke (Arch Systems), Abby Eon (Velotic), Bryce Barnes (LG CNS) and Sahil Yadav (AOI), moderated by Jeff Winter of IIoT World.
The starting numbers
According to the Manufacturers Alliance 2026 Great Acceleration research, 63% of manufacturers had to clean and map their data before launching AI pilots, and 21% cite inconsistent data formats across plants as their main obstacle. The panel argues, however, that data cleaning is not enough: what fails to travel is operational context.
Readiness is assessed per use case
A plant may be ready for downtime attribution yet lack the data needed for quality prediction. To check transferability, the panel proposes a test: can people from another site, who have never worked at the original one, launch the use case using only what is documented? If not, the pilot relied on the organization's tacit knowledge.
Four conditions are given for a use case to survive the transfer:
- the data must be linkable;
- operational context must exist beyond asset data;
- the data flow must reliably report failures at any hour;
- the output must fit into an existing workflow or system.
UNS and a data foundation are not the same thing
The panel distinguishes two layers. The Unified Namespace sits beneath MES, SCADA, APS and historians and deals with data accessibility and transport. A unified data foundation sits above these systems and harmonizes meaning: how the data relates to the production cycle and what state the machine was in when it was recorded. Confusing the two layers is cited as the most common mistake in AI projects; the UNS remains a necessary first step.
The silent failure of semantics
At the second site, tags have different names or subtly different meanings. The model does not stop with an obvious error: it degrades gradually, and operators lose trust before anyone identifies the cause.
Proposed countermeasures
- standardize data contracts and deployment models internally, without waiting for a standards body;
- attach provenance to every AI recommendation: which data produced it and which model version;
- capture relationships with an ontology and a graph model.
"The pipeline is the asset, not the model." (Sahil Yadav)
The context that resists automation
An example: an operator runs a machine at 70% utilization because spare parts will arrive in six weeks, and every dashboard flags it as underperforming. That judgment lives in the operator's head. For the panel, the operational context model should be treated as core infrastructure.
What this means for PLC, SCADA and integration work
For those working on PLC, SCADA and integration, the practical guidance is concrete: consistent tag naming and semantics across sites, explicit data contracts and provenance tracking. These elements, more than the model, allow a pilot to grow beyond a single plant. The transferability test is also useful as an internal check: if the documentation is not enough to replicate the use case elsewhere, the formalization work is not finished.
Caveats
The content is sponsored and the speakers are vendors, so the angle is inevitably commercial. The article also presents no real-case data on failures at the second plant: its claims should be read as good-practice guidance, not as quantitative evidence.